[Stener1]

Summary

Opening a tool call that returned a picture (a Read of a .png, an MCP screenshot tool) showed only [image]. It now shows the picture: a thumbnail (up to 320 px tall) where [image] stood, which opens at full size on a click.

The picture comes from the conversation's own record (the tool_result's base64 image block), not from the file. So it shows for files outside the worktree (/tmp/…), and for files changed or deleted since.

 GET …/sessions/{name}/transcript/tool/{id}
-  { output: "…\n[image]", … }
+  { output: "…\n[image]", images: [{ n: 0, line: 1, mediaType: "image/png" }], … }   # never the bytes
+GET …/sessions/{name}/transcript/tool/{id}/image/{n}            # the bytes, Content-Type image/*
+GET …/sessions/{name}/subagents/{id}/tool/{tool}/image/{n}      # the same for a helper's call

 <ToolRow>                      (conversation-view.tsx)
   <ToolDetailView image={…}>
-    <pre>{output}</pre>
+    outputParts(output, images)   (lib/tool-images.ts: each picture at its line)
+      <pre>{text}</pre>
+      <ToolImage n load name>     (tool-image.tsx: boxBlob → blob URL, fetched once the row opens)
+        <Dialog> full size

• Box (internal/transcript/detail.go): Detail lists the result's pictures in the same pass that reads its text, and skips their data. line is where each picture stands in output: its own [image] line, or the … more lines … line when a long output's middle is cut. So a text block that itself reads [image] stays text. Image decodes only picture n and stops reading at the result. It serves only PNG, JPEG, GIF and WebP (nosniff, a sandbox CSP, Cache-Control: private, immutable).
• App: fetches the pictures only once a row is opened, through boxBlob, because a plain <img src> can't carry the token. Anything the box can't give back (the record rotated past its 32 MB window, another type) shows [image] as before. It shows no toast for this.
• The chat's 2 s polling and the 32 KB output cap never carry image bytes.
• Codex sessions are unchanged, since their results hold no image blocks.

Evidence

The real app is driven against a dev box of this branch (.claude/skills/verify-berthd, new features/chat-tool-calls.md, scenario chat-tool-image-live). A real Claude Code reads a synthetic /tmp/acme-shot.png (480x300) from the chat.

• Before (main's berthd + main's app, same box): the Read opens to [image].

  

• After (terminal driver): the Read opens to the picture. Its naturalWidth is 480, and the row shows no [image] text.

  

• After: a click opens it at full size, and Escape closes it.

  

• After: /tmp/acme-shot.png is deleted, then the row is closed and opened again. The picture still shows, read from the record (screenshot).

• After (structured driver, where the chat is berthd's own host log): the same scenario passes (screenshot).

• Second view: the scenario fetches …/transcript/tool/{id}/image/0 through the agent. Its body is byte-for-byte the PNG Claude read, with Content-Type: image/png. The console shows no errors.

Tests:

internal/transcript  TestDetailNamesTheImagesAResultHoldsButNotTheirBytes
                       text + png + tiff + gif → images [{0,png},{2,gif}], output unchanged, no base64 in the JSON
                       Image(r1, 0|2) → the exact bytes; 1 (tiff), 3, -1, unknown id → ErrNoImage
                     TestAToolsImagesStandAtTheirOwnLines
                       a text block reading "[image]" before a real picture → the picture's line is its own
                       a 112 KB output with 3 pictures → head, "… more lines …", tail lines
internal/box         TestAToolsImagesAreServedFromTheTranscript
                       session and helper routes: image 0, 1 → bytes + Content-Type; 2, x, unknown call → 404
app/src/lib          tool-images.test.ts (placement, a literal [image] stays text, a cut output)
app/e2e/chat.spec.ts "a call that returned a picture shows it once opened, and opens it at full size"
                       not requested before the row opens; thumbnail and full size naturalWidth 640

Checks run locally (fork PRs wait for approval before CI runs):

• Passed: go vet ./..., go test -race ./..., pnpm check:*, pnpm test (632), pnpm build, npx tsc -p e2e, the touched e2e specs, and docs-site build.
• Failed, and fails on main too: TestSleepDurable (internal/box/runs) failed once under load and passed 5/5 when rerun. keyboard.spec.ts "⌘⇧E puts the keyboard in the Files panel" fails on origin/main as well.
• Not run: cargo check. This box lacks the GTK dev libraries, and no Rust changed.

Merge Danger

Door: two-way

A new GET route and an additive JSON field. Nothing is written anywhere, so a revert removes the change cleanly. An older app ignores images, and an older box sends none, so the app shows [image] as before.

Blast Radius: chat

• Only an opened tool row changes.
• The new route reads the same transcript window as the tool detail, and only for paired peers.
• The mock's release chat now has a third Read (checkout-result.png). "a tool call opens to what it did" looks for "Read 3 files".

🤖 Generated with Claude Code

[vercel]

@Stener1 is attempting to deploy a commit to the seanbrydonio's projects Team on Vercel.

A member of the Team first needs to authorize it.

[Stener1]

Opened as a draft because the author's ready-PR limit is full; it is marked ready automatically when a slot frees. I can't add the run-ci label from a fork: a maintainer please add it. The checks listed under Evidence were run locally.